Research On Topologies Of Echo State Network

With the rapid development of data processing calculations, data have more andmore complex, including the highly non-linear, chaotic, noisy large. Computationalproblems of this nature, the traditional methods of mathematical and neural networks,such as calculus, differential, artificial neural network, traditional recurrent neuralnetworks, have been gradually incompetent. However, echo state network (ESN) whichhas the characteristics of simple training process and global-optimization gainwidespread concern.For its excellent computing power, especially for the computing power oftiming-related data, ESN used to predict, control, system identification, speechrecognition, anomaly detection, medical diagnostics and other fields. Accordingly,considering the influence of reservoir structure on the computing power of ESN, thisdissertation focuses on reservoir network topologies of ESN, and do some experiment.In this paper, the classification ESN will be inspired by complex networktopologies, such as small-world and scale-free topologies and a representative regularnetwork. Compared with random topology, the method is tested on the benchmarkprediction problem of Nonlinear Autoregressive Moving Average system (NARMA),Mackey-Glass time series and Lorenz system in MATLAB simulation platform.The result shows that the ESN with small-world topology demonstrate strongstability and high accuracy, when tested on NARMA, Mackey-Glass time series andLorenz system.